Copyright concerns remain my main reason for not using AI for programming
Disclaimer: I am not a lawyer, and this is not meant as legal advice!
I have a number of concerns regarding using generative-AI tools: the risk of cognitive atrophy; not wanting to spend time correcting mistakes within automated output; potential cost increases as VC subsidies run out; and so on. However, copyright concerns are probably my number one reason for staying away from these tools.
It seems that using AI for programming puts you in a double bind. On one hand, AI-generated code (like other AI-generated output) https://www.congress.gov/crs-product/LSB10922. This means that, whenever programmers state that they (or their company) created a project entirely via vibe coding, they're essentially saying that that code is in the public domain should it get leaked. ( https://fortune.com/2026/03/31/anthropic-source-code-claude-code-data-leak-second-security-lapse-days-after-accidentally-revealing-mythos/.)
On the other hand, https://arxiv.org/pdf/2508.16853 that a given set of gen-AI-created code will contain enough copyrighted material to either infringe on a proprietary copyright or force you to release your source code (at least in some cases) under a copyleft license like the GPL. This could result in monetary damages or (perhaps worse yet for some companies) force proprietary code to be released under an open-source license.
I see a few potential ways around this problem:
1. Treat all code produced by an LLM as if it falls under a proprietary or copyleft license. In other words, you can incorporate the idea or method expressed in the code into your own project, since https://www.copyright.gov/circs/circ33.pdf, but you should avoid copying the code itself into your project unless (A) https://www.copyright.gov/comp3/chap300/chap300-draft-3-15-19.pdf or (B) https://www.copyright.gov/fair-use/. This is already my approach for StackOverflow code, which is released under a (copyleft) CC-BY-SA license.)
2. As suggested by https://arxiv.org/pdf/2508.16853, use an LLM that has only been trained on public-domain or permissively-licensed code. (Permissive licenses, unlike copyleft ones, don't require that you release your own code under the same license.) In addition, this LLM would need to inform you when enough code from a given source was used that you'd need to provide attribution to the copyright owner. (I'm not aware of any easily-accessible LLM that meets these requirements, but if you are, please do let me know.)
3. Don't use LLMs. This way, you can check the license of all code that you're referencing for a given project and determine exactly how to apply this code within your own work.
(Some might offer a fourth solution: Use LLMs that come with copyright indemnification protection, thus shielding you from copyright lawsuits. However, I would recommend reading their terms of service very, very carefully. For instance, under Anthropic's Commercial Terms of Service, we read:
"Additionally, Anthropic’s defense and indemnification obligations will not apply to the extent the Customer Claim arises from: (a) modifications made by Customer to the Services or Outputs; (b) the combination of the Services or Outputs with technology or content not provided by Anthropic; (c) Inputs or other data provided by Customer;"
Again, I'm not a lawyer, but I'd interpret this to mean that once I modify the output of AI-generated code (which I imagine to be a pretty routine task), I may lose my indemnification protection for that part of my codebase.)
TL;DR: I think copyright concerns are often overlooked when it comes to LLM output--and not something that can be solved simply with more powerful, advanced models. So I'll keep avoiding
